⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡀⠀⠀⢀⣤⣀⠀⠀⢀⣴⢶⣄⠀⠀⣀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣀⣀⠀⠀⣠⡾⠛⢷⣄⣿⠁⠉⠻⠾⠛⠁⠀⠙⢛⣛⠻⣧⣤⣄⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣿⠉⠉⠛⠛⠋⠀⢀⣀⣉⢀⣠⣤⣴⠛⠲⣤⡖⠋⠉⠈⠙⣦⣦⣿⣇⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣠⡴⠟⠀⠀⣀⣀⣀⣴⣋⡀⠉⠉⠀⠀⣹⡶⢟⠛⠛⡛⢶⣄⣠⣿⣇⡀⠈⠙⡷⣆⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⣴⠟⢻⣤⣤⡶⠟⠛⠛⡙⠛⢻⣿⣿⣶⣤⣾⣯⣤⠾⠷⠶⠷⠶⠿⠿⣯⣬⣻⣦⣼⡷⠟⢻⡆⠀⠀
⠀⠀⠀⠀⠀⠀⢀⣴⠾⠃⢀⣴⠟⠉⠀⠰⠄⢀⡧⠤⣞⣠⡾⠟⠋⠉⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠉⠛⢷⣄⡀⠘⣷⡀⠀
⠀⠀⠀⠀⠀⠀⣿⠁⠀⢀⣿⠏⠀⠀⠃⣠⣴⣾⣦⡾⠋⠁⠀⠀⠀⠀⠀⠀⣠⡤⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⠉⢿⡆
⠀⠀⠀⠀⣠⡾⠋⠀⠀⣸⣿⡀⠘⠃⠸⢿⣿⠟⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠹⠷⠀⠀⠀⠀⠀⣾⡃⠀⠀⠀⠀⠀⠙⣧⠙⣷
⠀⠀⠀⠀⣿⠀⠀⠀⠀⣻⣿⣿⣿⣷⣿⡿⠋⠀⠀⠀⠀⠀⢀⣶⠂⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠋⠀⠀⠀⠀⠀⠀⠸⣿⠋
⠀⠀⠀⠀⣼⠷⢶⣴⡾⠿⣿⡏⠀⣴⠏⠀⠀⠀⣤⠀⠀⠀⠈⠻⠆⠀⠀⠀⠀⠀⢸⣇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢿⠀
⠀⠀⠀⣸⠏⣰⡿⠋⢦⣠⡼⢇⣼⠃⠀⠀⠀⠀⠛⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢠⣿⠂⠀⠀⠀⠀⣼⠀
⠀⠀⣰⠏⣰⡟⠀⠠⢀⣯⣀⣾⠃⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣤⡄⠀⠀⠀⠀⠀⠀⠀⠀⠙⠁⠀⠀⠀⢠⡟⠀
⠀⢠⣟⣠⣿⠀⠰⠀⣿⣿⣿⠃⠀⠀⠀⠀⠀⣼⠃⠀⠀⠀⠀⠀⠀⠀⠀⠀⠻⠆⠀⠀⠀⠀⠀⠀⠀⠀⢀⣀⣤⡴⠾⠋⠀⠀
⠀⣾⣹⠉⣿⣦⣤⠴⠒⣿⠃⠀⠀⠀⠀⠀⠀⠻⠖⠀⠀⠀⠀⣷⡀⠀⠀⠀⠀⠀⠀⠀⢀⣀⣤⡴⠶⠛⠋⠉⠀⠀⠀⠀⠀⠀
⢰⡟⣏⠀⢻⣟⠁⣀⣼⠇⠀⠀⢀⡄⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣀⣤⡴⠶⠛⠛⠉⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⣾⢷⡋⠇⠈⢿⣦⣰⠏⠀⠀⠀⠘⠧⠀⠀⠀⠀⠀⢀⣀⣠⣤⠶⠞⠛⠉⠉⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⣿⢸⣇⠸⢢⠀⣹⡟⠀⠀⠀⠀⠀⢀⣀⣤⡴⠶⠛⠋⠉⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⢻⣟⢘⡆⢈⣴⠏⢀⣀⣤⡴⠶⠛⠋⠉⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠙⠷⠶⠿⠷⠛⠋⠉⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀

████████╗  █████╗  ██████╗ ██████╗       ██╗  ██╗██╗████████╗
╚══██╔══╝ ██╔══██╗██╔════╝██╔═══██╗      ██║ ██╔╝██║╚══██╔══╝
   ██║   ███████║██║     ██║   ██║█████╗█████╔╝ ██║   ██║   
   ██║   ██╔══██║██║     ██║   ██║╚════╝██╔═██╗ ██║   ██║   
   ██║   ██║  ██║╚██████╗╚██████╔╝      ██║  ██╗██║   ██║   
   ╚═╝   ╚═╝  ╚═╝ ╚═════╝ ╚═════╝       ╚═╝  ╚═╝╚═╝   ╚═╝   

🌮 taco-kit

Research → Plan → Implement → Verify + PR Review

Interactive agents that ask you questions before starting work — front-loading the context you'd otherwise reveal across 5 follow-up correction turns.

Installation

# Install globally (recommended)
uv tool install git+https://github.com/rosstaco/taco-kit

# Or run without installing
uvx git+https://github.com/rosstaco/taco-kit init

Install agents into a repo

# Install to the current repo (fetches latest from GitHub)
taco-kit init

# Install from a local checkout
taco-kit init --source /path/to/taco-kit

# Overwrite existing files without prompting
taco-kit init --force

This creates .taco-kit/agents/ and .taco-kit/conventions.md in your repo, and updates .vscode/settings.json so VS Code discovers the agents automatically.

Install agents to your Copilot CLI profile

taco-kit init --profile

This writes agent files to ~/.copilot/agents/ so they are available in every project via the Copilot CLI:

copilot --agent=01-Research
# or use /agent in interactive mode

The Rules

1. Each phase produces a concrete artifact — not instructions about what to do, but the actual output.
2. Each phase gates the next — Plan requires Research output. Implement requires Plan output. Verify requires running code.
3. Requirements go in the prompt, not in follow-up corrections — if you know what "done" looks like, say it upfront.
4. One session per task — don't open new sessions for the same prompt. Fix the prompt, re-run.
5. Follow project standards — look for coding standards, conventions, or architecture docs in the repo. If none exist, match the existing code patterns.

Structure

README.md                          ← This file
.taco-kit/
├── conventions.md                 ← Shared rules: plan storage, project standards
└── agents/
    ├── 01-research.agent.md       ← Gathers real data, schemas, constraints
    ├── 02-plan.agent.md           ← Architecture, test cases, acceptance criteria
    ├── 03-implement.agent.md      ← TDD build with E2E verification
    ├── 04-verify.agent.md         ← Independent review with evidence
    └── pr-review.agent.md         ← Structured PR review (standalone)

Plan artifacts are stored per feature:

project/plans/YYYY-MM-DD-feature-name/
├── research-findings.md
├── plan.md
└── verify-report.md

Agents

Phase	Agent	Questions Asked	Output
Research	01-research.agent.md	4 — problem, data sources, codebase, constraints	research-findings.md
Plan	02-plan.agent.md	5 — requirements, display, errors, done-criteria, gotchas	plan.md
Implement	03-implement.agent.md	3 — confirm plan, verification method, pitfalls	Working code + E2E proof
Verify	04-verify.agent.md	3 — verify against, what changed, concerns	Verification report
PR Review	pr-review.agent.md	4 — PR number, plan/issue, focus, concerns	Structured review

Usage in VS Code

After running taco-kit init, agents appear in VS Code Chat automatically. Reference them with:

@.taco-kit/agents/01-research.agent.md
@.taco-kit/agents/02-plan.agent.md
@.taco-kit/agents/03-implement.agent.md
@.taco-kit/agents/04-verify.agent.md
@.taco-kit/agents/pr-review.agent.md

Usage in Copilot CLI

After running taco-kit init --profile, use agents in the CLI:

# Interactive — pick from agent list
copilot
/agent

# Direct — specify agent by name
copilot --agent=01-Research -p "Research the GitHub API"

CLI Reference

taco-kit init [OPTIONS]

Options:
  --profile       Install to ~/.copilot/agents/ for global CLI use
  --path, -p      Target repo directory (default: current)
  --force, -f     Overwrite existing files without prompting
  --source, -s    Load files from a local path instead of GitHub
  --help          Show help

Development

git clone https://github.com/rosstaco/taco-kit.git
cd taco-kit
uv sync
uv run pytest -v
uv run taco-kit init --source . --path /tmp/test --force

Multi-Agent Reviews

When running parallel agents, always specialise by concern — not by model. See examples in each agent file.

License

MIT